List your company

Long-context evaluation on RULER 128k

98Query Metric (MQ)Llama-3.1-8B
-2.8823.3149.575.69May 26, 2025Jul 7, 2025Aug 19, 2025Sep 30, 2025Nov 12, 2025Dec 24, 2025Feb 5, 2026

Evaluation results

MethodLinks
Llama-3.1-8BBackbone=Llama-3.1-8B, KV Cache Budget=Full2025.059810010010010088649695.21.666.6643677.6—
TailorKV-1Backbone=Llama-3.1-8B, KV Cache Budget=128+(896) tokens2025.059892921001006409367.20.416604063.3—
TailorKV-2Backbone=Llama-3.1-8B, KV Cache Budget=128+(896) tokens2025.05971009210010064169685.60.440643668.5—
Yi-9BBackbone=Yi-9B, KV Cache Budget=Full2025.05841001001009680286910.43.689.3363664—
TailorKV-2Backbone=Yi-9B, KV Cache Budget=128+(896) tokens2025.0580100100969676205548.82441.3363261.9—
TailorKV-1Backbone=Yi-9B, KV Cache Budget=128+(896) tokens2025.0579.6100100969672204419.223.244403258.9—
QuestBackbone=Llama-3.1-8B, KV Cache Budget=1024 tokens2025.057180680884806659.20.452482846.8—
Yi-6BBackbone=Yi-6B, KV Cache Budget=Full2025.056710010010084724304.81.2100322455.3—
TailorKV-2Backbone=Yi-6B, KV Cache Budget=128+(896) tokens2025.0548100100100842842114.41.260322047.1—
TailorKV-1Backbone=Yi-6B, KV Cache Budget=128+(896) tokens2025.054710010010084160253.20.861.33222.745.5—
Qwen3-8B-SFT w/ LongRModel=Qwen3, Parameter Size=8B, Training Strategy=SFT w/ LongR2026.0245.59———————80.2————61.4758.63
SnapKVBackbone=Llama-3.1-8B, KV Cache Budget=1024 tokens2025.0538100840842401965.6028483240.2—
Qwen3-4B-SFT w/ LongRModel=Qwen3, Parameter Size=4B, Training Strategy=SFT w/ LongR2026.0235———————90.84————60.7156.28
Qwen3-8B-SFTModel=Qwen3, Parameter Size=8B, Training Strategy=SFT2026.0230.5———————51.96————48.0361.64
Qwen3-8B-SFT w/ DAPOModel=Qwen3, Parameter Size=8B, Training Strategy=SFT w/ DAPO2026.0230.4———————76.56————54.8757.64
Qwen3-4B-SFT w/ DAPOModel=Qwen3, Parameter Size=4B, Training Strategy=SFT w/ DAPO2026.0228.7———————90.48————57.9254.57
QuestBackbone=Yi-9B, KV Cache Budget=1024 tokens2025.0528100844722402816.80.869.3243237.1—
Qwen3-4B-SFTModel=Qwen3, Parameter Size=4B, Training Strategy=SFT2026.0227.2———————91.44————57.2253.02
QuestBackbone=Yi-6B, KV Cache Budget=1024 tokens2025.05239672072200176.40.849.316829.2—
Qwen3-8B-BaseModel=Qwen3, Parameter Size=8B, Training Strategy=Base2026.0217.6———————47.68————39.0551.88
Qwen3-4B-BaseModel=Qwen3, Parameter Size=4B, Training Strategy=Base2026.0214———————45.64————35.4646.74
PQCacheBackbone=Yi-9B, KV Cache Budget=1024 tokens2025.05581602440240.477.3162814.2—
StreamLLMBackbone=Llama-3.1-8B, KV Cache Budget=1024 tokens2025.054040040500.49.324205.4—
SnapKVBackbone=Yi-9B, KV Cache Budget=1024 tokens2025.054921202040127.2253.3203219.9—
SnapKVBackbone=Yi-6B, KV Cache Budget=1024 tokens2025.05476001600180.869.3201616.2—
PQCacheBackbone=Llama-3.1-8B, KV Cache Budget=1024 tokens2025.05308048020.8066.6403212.6—
PQCacheBackbone=Yi-6B, KV Cache Budget=1024 tokens2025.052080124010054.6242810.2—
StreamLLMBackbone=Yi-9B, KV Cache Budget=1024 tokens2025.051.1404400020086.6162410.6—
StreamLLMBackbone=Yi-6B, KV Cache Budget=1024 tokens2025.051044000200.86820168.9—